Fix ensure_positive_ndt for models with trial-to-trial ndt variability - #1292
Open
EItanm1999 wants to merge 1 commit into
Open
Fix ensure_positive_ndt for models with trial-to-trial ndt variability#1292EItanm1999 wants to merge 1 commit into
EItanm1999 wants to merge 1 commit into
Conversation
Contributor
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Team Run ID: 📒 Files selected for processing (3)
Included review availability: Your plan provides up to 2 included reviews per hour; 1 remains after this review. 📝 WalkthroughWalkthrough
ChangesNon-decision time support boundary
Estimated code review effort: 2 (Simple) | ~10 minutes Merge Risk: ⚪ Minimal · up to This localized change corrects response-time support handling for models with trial-level non-decision-time variability and adds focused regression coverage; no actionable merge-blocking risk remains beyond normal checks and review. Suggested reviewers: 🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
Full details: Docstring CoverageExplanation Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 5 functions across 2 files. (1 skipped: 1 unsupported.) ✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
ensure_positive_ndt floors the log-likelihood wherever rt - t <= 1e-15. That is the correct support edge for a fixed non-decision time, but not for any model carrying st: there t is drawn per trial from Uniform(t - st, t + st), so the fastest admissible response time is t - st and the band [t - st, t] carries real density. The guard replaced that entire band with LOGP_LB regardless of what the likelihood returned. Applied unconditionally at both call sites, affecting every *_st model. Measured on ddm_st: compiling HSSM's observed-RV logp for a model whose likelihood IS the exact ddm_st quadrature and comparing against the same quadrature called directly gave differences of -5.3 to -12586.8 nats, varying with theta. With the guard skipped the two agree to exactly 0.0000 on 7 of 8 parameter vectors (the 8th had a subject z outside its bound). So this guard accounts for the entire discrepancy while parameters stay in bounds. p_outlier partially masks it: the floored value is wrapped in the lapse mixture, so affected trials emerge at log(0.05/20) = -5.99 rather than at LOGP_LB, which is why this presented as bad geometry rather than an obvious -inf. The t - st edge is stated as correct under the half-width convention of ssms/cssm, which every LAN *_st model follows, rather than as a universal law. full_ddm is the one bundled model on the other convention: its only likelihood is the blackbox wrapper around hddm_wfpt, which reads st as a full width and puts its own edge at t - st/2. No separate factor is needed, because that edge sits above t - st and hddm_wfpt already returns zero density across [t - st, t - st/2), which the wrapper maps to the same lower bound. Verified: across a 0.28-0.46 RT scan the raw likelihood's highest floored rt is exactly t - st/2, and the guard changes no value at any rt (max |delta| 0.0). Only st moves the response-time support edge. sz (starting point) and sv (drift) do not, and are deliberately not consulted; a regression test pins that down. Tests are one parametrized case per convention (fixed t, st moves the edge, sz/sv leave it) over an explicit RT vector that straddles all three edges, so the in-band assertion no longer depends on an unseeded draw. They compare against the imported LOGP_LB rather than a -66.1 literal: under PYTENSOR_FLAGS=floatX=float32 the bound is -66.0999984741211, and wrapping it in np.array defeats NumPy's weak promotion, so the literal form failed there. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
EItanm1999
force-pushed
the
fix/ensure-positive-ndt-st-aware
branch
from
September 2, 2026 20:58
3834b46 to
25a487f
Compare
Author
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
ensure_positive_ndt floors the log-likelihood wherever rt - t <= 1e-15. That is the correct support edge for a fixed non-decision time, but not for any model carrying st: there t is drawn per trial from Uniform(t - st, t + st), so the fastest admissible response time is t - st and the band [t - st, t] carries real density. The guard replaces that entire band with LOGP_LB regardless of what the likelihood returned.
The comment above the call site states the assumption ("Assuming that the non-decision time parameter is always named t") but nothing checks whether the model also has st. Affects every *_st model, applied unconditionally at dist.py:607 and :617.
Measured on ddm_st: compiling HSSM's observed-RV logp for a model whose likelihood IS the exact ddm_st quadrature and comparing against the same quadrature called directly gave differences of -5.3 to -12586.8 nats, varying with theta. With the guard skipped the two agree to exactly 0.0000 on 7 of 8 parameter vectors (the 8th had a subject z outside its bound). So this guard accounts for the entire discrepancy while parameters stay in bounds.
p_outlier partially masks it: the floored value is wrapped in the lapse mixture, so affected trials emerge at log(0.05/20) = -5.99 rather than at LOGP_LB, which is why this presented as bad geometry rather than an obvious -inf.
Only st moves the response-time support edge. sz (starting point) and sv (drift) do not, and are deliberately not consulted; a regression test pins that down.
Summary by CodeRabbit
Bug Fixes
Tests
Documentation